Make Multitext an actual Mapping, breaking keys(), len(), and duplicate langs

[imnasnainaec]

Multitext hand-wrote get, __contains__, keys, values and items on top
of __getitem__ / __iter__ / __len__ — the set collections.abc.Mapping
derives from those three. Inheriting it drops the five, and makes
isinstance(mt, Mapping) true.

What changes for a caller

	before	after
keys(), values(), items()	lists	views
len(mt)	every form, lang-less included	languages
a language on two forms	two keys; values() had both texts	one key, the first form; forms keeps both
a lang-less form	mt[None] returned its text, None in mt was False	not a key at all
del mt["en"], two en forms	dropped the first, "en" in mt stayed true	drops both
repr(mt)	always dict-shaped	pairs when the forms are not one per language

Views

Typing the class as a Mapping while returning lists is three suppressed Liskov
violations: a list is not a KeysView.

• A caller typed against Mapping[str, Text] would be promised set operations it
  has not got. Those work now, and a caller wanting a list can say so.
• Cost: one assertion in the suite, which mypy pointed at, and one line in the
  guides.

One key per language, and len() counting them

A choice, not something the ABC forces: multidict.MultiDict is a registered
Mapping whose keys() repeats and whose len() counts pairs — what this class
did.

• Deduping is what gets __getitem__, len(), dict() and the three views to
  agree with each other.
• Before, len(mt) counted a lang-less form that keys() had always excluded, so
  it could exceed len(mt.keys()); and a repeated language made len(mt) 2 where
  dict(mt) held 1 key.
• werkzeug's MultiDict makes the same call for the same reason.

A lang-less form is not a key

The internal lookup matched it, so mt[None], mt.get(None) and
None in mt.keys() all reached it while __iter__, the views and len() left it
out. That is a key the mapping answered for but never reported — and a KeysView,
which is a Set, claiming an element it would not yield.

Deletion by language, assignment by form

del mt["en"] takes every en form, so the key is gone afterwards. Assignment
updates only the form it names.

• Both MultiDicts collapse on assignment instead — safe for them, because a
  value is a str.
• Here the mapping hands back a Text, but the Form carrying it also holds
  annotations and out-of-schema residue that no key reaches.
• This is the wrong library for the most-used mutator to be the one operation that
  silently discards residue. forms stays the way to edit a duplicate
  deliberately.

Repr shape

A repeated or lang-less form rendered as a dict literal that cannot exist, whose
keys contradict keys(). The inherited views repr through the mapping, so that
text reached KeysView(...) too.

Multitext({'en': 'dog', 'fr': 'chien'})          # one form per language
Multitext([('en', 'colour'), ('en', 'color')])   # a repeated language
Multitext([(None, 'x')])                         # a lang-less form

• Dict-shaped is kept for forms that are one per language — all a schema-valid
  document has.
• The pair-list fallback keeps every form visible, with the shape itself as the
  signal that forms holds more than the mapping reaches.

What does not change

• Fidelity, at all. The writer emits every form off forms, duplicates and
  lang-less ones included, and consults only __bool__; no writer path touches
  len() or the views. duplicate-form-lang reads forms too.
• bool(mt) still answers "is there anything to serialize", which residue and a
  lang-less form each defeat on their own, so it stays independent of len().
• MutableMapping is still not inherited. clear and popitem have no clear
  meaning for a form list that can hold forms no key reaches, so the two mutators
  the class already had are still the only two.

How rare these shapes are

Over the 34,904 multitexts reachable from an entry in the 24 loadable 0.13 corpus
files:

• duplicate languages: 0 — they appear only in the hand-authored negative
  fixture
• lang-less forms: 2, both in the LIFT spec's own example documents

So this is about the mapping being coherent for input the reader accepts, not
about a shape most files carry.

Coverage

• tests/test_text.py, new — the mapping semantics, which need no reader.
  • The views, insertion order, first-form resolution, both mutators, deleting
    through the mapping while iterating it, truthiness, equality, both repr shapes.
• tests/test_reader.py — the reader's share, on the two negative fixtures that
  already carry these shapes.
  • Replaces a hand-built model that had no business in a reader test.
• tests/test_property_roundtrip.py — the Multitext strategy drew unique,
  never-None langs, so the round-trip properties never saw either shape.
  • Langs now come from the same pool plus None. Over 600 draws, roughly a third
    of multitexts carry a duplicate and a fifth a lang-less form, with most
    examples still schema-valid.
• tests/test_writer.py — a multitext holding only residue and no forms.
  • Empty as a mapping, so the writer's decision to emit it cannot come from
    len(). No corpus fixture has that shape, and losing it would drop the
    attribute silently.

Guide and CHANGELOG updated to match.

python scripts/check.py green: 586 passed, 98% coverage. mkdocs build --strict
green.

🤖 Generated with Claude Code

---

This change is